NVIDIA · NCP-GENL

The adaptation-strategy thread

Choosing how to change — or not change — a model's behavior. Opens in M1 with the architecture families a task has to match, runs through M2's prompting-vs-RAG-vs-fine-tuning decision rule, and closes in M5 where LoRA/PEFT and the SFT/RLHF/DPO/GRPO alignment family are the concrete mechanisms behind that decision.

NCPG-T2 · 15 lessons across 3 modules

  1. M1M1-01Scaled dot-product attention: Query, Key, Value, and the √dₖ scale
  2. M1M1-02Multi-head attention, positional encoding, and layer normalization
  3. M1M1-03Architecture families: encoder-only, decoder-only, and encoder-decoder
  4. M1M1-04Embeddings: extraction from encoder and decoder models, and cosine similarity
  5. M1M1-05Output sampling for decoders: greedy, beam search, temperature, top-k, top-p
  6. M2M2-01In-context learning: zero-shot, one-shot, and few-shot prompting
  7. M2M2-02Chain-of-thought prompting and prompt templates for small or specialized data
  8. M2M2-03Causal language modeling: the training objective behind generation
  9. M2M2-04Output control: constrained decoding and validation wrappers
  10. M2M2-05Choosing prompting vs. RAG vs. fine-tuning under a stated constraint
  11. M5M5-01Parameter-efficient fine-tuning: LoRA, adapters, and P-tuning
  12. M5M5-02Alignment: SFT, RLHF, DPO, and GRPO
  13. M5M5-03Contrastive loss for embeddings
  14. M5M5-04Early stopping and fine-tuning impact assessment
  15. M5M5-05When to fine-tune vs. prompt or RAG

Part of the throughlines running across the NCP-GENL prep course.